I.
Editing what's already in the image
Most of my present-day work is on teaching diffusion and autoregressive image models to change what's already there in an image — swap a label, fix a typo, alter a price — without redrawing everything around it. It's harder than it sounds: type and graphic context have a font, a weight, an alignment, a shadow, and these must survive the edit unchanged. The method is called IDAttn and the paper is going to ICML, on multi-instance editing in general, with infographics as the central application.
